BN for Language Identification

BN for Language Identification

1. introduction of BN and S(tack)BN

2. computation of BN

如下图，首先通过DCT->Bark filtering->FDLP(得到包络)->compression(静态压缩是对前一步的结果取log，符合人耳的听觉感受，动态压缩是能够体现前一步结果的动态变化)->nomalization->first NN
然后第一个网络结束计算之后，bottleneck(80 dimensions)->stack to 400dimensions->归一化->second NN